Development of real-time Deep Reinforcement Learning (DRL) Hardware

The Army Research Laboratory Research Associateship Program (ARL-RAP) offers a research opportunity focused on developing real-time hardware for Deep Reinforcement Learning algorithms in tactical applications. Current approaches optimize machine learning training by exploiting Deep Neural Networks sparsity with compute-intensive floating-point 32-bit representation for non-zero valued network parameters. This research aims to improve these approaches and incorporate High Level Synthesis (HLS) to obtain hardware designs optimized for various criteria including power, latency, and computation. The ARL-RAP program is designed to significantly increase the involvement of creative and highly trained scientists and engineers from academia and industry in scientific and technical areas of interest and relevance to the Army. The Computational and Information Sciences Directorate (CISD) conducts research in disciplines relevant to achieving the digital battlefield, focusing on sensing, distribution, analysis, and display of information in modern battle spaces. Research at ARL focuses on communications, atmospheric modeling, battlefield visualization, and computing.